使用Paddle OCR实现并行处理失败问题求助

问题现象 #

• 单线程（num_threads=1）调用predictx_parallel方法时，OCR识别正常运行；
• 多线程（num_threads>1）调用时，触发两类框架层面错误：

错误信息 #

UnimplementedError: Currently, only can set dims from DenseTensor or SelectedRows. (at /paddle/paddle/fluid/framework/infershape_utils.cc:314)
      [operator < fused_conv2d > error]

NotFoundError: Variable Id 29797 is not registered.
      [Hint: Expected it != Instance().id_to_type_map_.end(), but received it == Instance().id_to_type_map_.end().] (at /paddle/paddle/fluid/framework/var_type_traits.cc:103)
      [operator < fused_conv2d > error]

原实现代码 #

def predictx_parallel(input_images: List[Image], ocr_params: OcrParams, num_threads: int) -> Tuple[List[Dict], List[Image]]:
    def ocr_image(image):       
        image_array = np.array(image)   # type(image) : <class 'PIL.Image.Image'>
                                        # type(image_array) : <class 'numpy.ndarray'> 
        results = ocr.ocr(image_array)  # [[[[[381.0, 285.0], [537.0, 285.0], [537.0, 333.0], [381.0, 333.0]], ('PAGE1A', 0.9997838139533997)], [[[388.0, 371.0], [530.0, 371.0], [530.0, 419.0], [388.0, 419.0]], ('PAGE1B', 0.998117983341217)]]]
        return results

    ocr = get_ocr_obj(params=ocr_params)  # <class 'paddleocr.paddleocr.PaddleOCR'>
   
    with ThreadPoolExecutor(max_workers=num_threads) as executor:
        results = list(executor.map(ocr_image, input_images))

使用环境 #

• paddlepaddle==2.6.0
• paddleocr==2.7.0.3
• python==3.9.12

问题原因 #

单个PaddleOCR实例无法在多线程环境下安全复用，跨线程共享会导致Paddle内部张量、变量管理出现线程安全问题，触发框架报错。

解决方案 #

改用多进程实现并行处理，在每个进程内独立创建新的PaddleOCR实例，避免跨线程共享OCR对象。

修复后代码 #

pipeline.py

def predictx_parallel_processes(input_images, num_processes):
    with Pool(processes=num_processes) as pool:
        pool.map(ocr_image_x, input_images)

ocr_processing.py

def ocr_image_x(image):
    process_pid = os.getpid()
    logger.info(f"Process PID: {process_pid}")
    ocr = PaddleOCR()  # 每个进程创建独立的OCR实例
    image_array = np.array(image)
    results = ocr.ocr(image_array)
    logger.info(results)   

内容的提问来源于stack exchange，提问作者Soumya